This project is developed to demonstrate **DCGAN** network architecture which generates abstract art images.
1. Introduction
2. Architecture
4. Referecnces
Thsi project has implemented DCGAN network architecture which contains Generator and Discriminator model to generate sample of size (512 _ 512 _ 3) based on given training data.
It is composed of Generator and Discriminator model, To understand further..
The Generator model implements transpose convoluational layer to perform convoluation as well as upsample the output of convoluation using strides of (2,2) whicle keeping the padding same and applying leakyRelu activation to each genrated output. Graphical representation is as follows:
The Discriminator model implements convoluation layer to perform convoluation and apply maxpooling of (2,2) to downsample images and afterwards uses fullyconnected layer to output prediction by applying sigmoid activation. Graphical representation is as follows:
Following parameter were used while training the network
Image height : 512
Image width : 512
Image Channel : 3
batch size : 16
latent dim : 200
number of epochs : 200
network seems to generate close to real images out of provided latent dimension which can be seen during the training.
some of the sample that were generated by network during testing,
further model can be trained for more epoches to get better result from the network.
dataset link
other useful links
https://github.com/nikhilroxtomar/DCGAN-on-Anime-Faces/blob/master/gan.py